Latest AI news and analysis
Zero Retention Moves The Safety Alert To You
Anthropic's Enterprise Frontier Safeguards keeps the logs in your cloud account and sends the misuse flags to your team, with no vendor human review. The trade-off was not removed. It was transferred.
Read article
The Protocol Moves the Message. It Does Not Carry Your Authority.
On 13 August 2026 the Agentic AI Foundation announced 57 new members including Visa and Wells Fargo, taking it to 247. A June gap analysis of five agent interoperability protocols found that voting and dissent preservation are absent from every one of them, and concluded that governance is a missing layer above the standards rather than a missing feature inside them.
Read article
Your Agent Browses as You. Nobody Issued It a Credential.
Cloudflare shipped an agent-first browser on 6 August 2026, still in beta, whose defining design choice is that every session starts fresh. That is the exact opposite of the consumer agentic browsers, whose value comes from running inside the profile you are already logged into. Published security research shows what that inheritance costs, and the control is identity rather than filtering.
Read article
Claude's Text Now Carries a Watermark. It Marks the Model, Not the Author.
Anthropic now names Fable 5.1 and Mythos 5.1 as carrying Claude's text watermark, and has opened a detection API in private preview to organisations obligated under EU law. The part nobody is discussing is that both Anthropic and Google DeepMind say this class of mark is sparser on factual, constrained writing. That is exactly what regulated Australian work produces, which makes the mark useless as an internal control in both directions.
Read article
MCP Went Stateless. Your Audit Trail Just Changed.
The Model Context Protocol revision dated 28 July 2026 removes protocol-level sessions and the initialisation handshake. Cross-call state becomes explicit tool arguments, identity rides on every request, and one deleted feature quietly creates a duplicate-execution risk worth knowing about before your next agent integration.
Read article
Your AI Vendor Is Now an Identity Provider
Sign in with ChatGPT began rolling out across partner products on 29 July 2026, and OpenAI's own documentation says it is enabled by default for any organisation that has not set an explicit policy. The list of parties who can issue a credential into your systems just grew, and most access reviews do not have a column for it.
Read article
Your Agent Had a Critical CVE. There Was Nothing to Patch.
On 6 August 2026 Microsoft disclosed two critical vulnerabilities in cloud AI agent services, one scored 9.9. Both were already fixed, and both told customers there was nothing to do. That disposition quietly removes managed agents from vulnerability management and moves them into vendor risk, where most control environments are not looking.
Read article
Delegate by Task Length, Not by Task Type
METR measures AI capability as the length of task a model finishes at a given success rate. Read properly, it is the most useful delegation tool available, and it says something uncomfortable about regulated work.
Read article
One Serious Claim in Nine Is Now Mental Stress. A Decade Ago It Was One in Seventeen.
We computed the ten-year mental stress series from Safe Work Australia's national claims dataset. Serious mental stress claims rose from 6,261 to 16,839, their share of all serious claims nearly doubled, and each one costs around five times the median time lost. The full method is published with the finding.
Read article
Inkling's Weights Are Open. Its Model Family Tree Still Needs Evidence.
Thinking Machines released Inkling with downloadable weights, an Apache 2.0 licence and a disclosure that other models helped generate early training data. Open weights improve checkpoint control. They do not give procurement a complete model family tree. Build one with evidence states instead of an openness label.
Read article
You Requested Claude Opus 5. Check Which Model Answered.
An Opus 5 request can produce an Opus 4.8 answer after a safety-classifier decline. The requested model is configuration; the served model is evidence. Here are the fallback mechanics, the contamination rate that exposes mixed evaluations, and the evidence record that keeps attribution defensible.
Read article
Copilot Can See Your Screen. The Share Picker Is the First Control.
Microsoft 365 Copilot Vision lets a licensed user share an entire screen, a specific window or a mobile camera view during a voice conversation. For regulated Australian teams, the share picker now needs a purpose, a boundary and a stop condition before anyone starts talking.
Read article
A Phishing Link Built a Rogue Agent. Detect the Birth, Not Just the Run.
Zenity says one malicious link could once create, connect, publish and schedule a Workspace Agent inside a signed-in user's session, and that OpenAI fixed the reported flaw. The lasting control is to reconcile every agent's authorised birth record against its live security state.
Read article
OpenAI Presence Arrived. Decide Who Can Change the Agent.
OpenAI Presence makes post-launch agent improvement part of a managed service. That puts vendor engineers, integrators and customer teams inside one change loop. The contract must specify who may propose, test, approve, release, pause and reverse each production change.
Read article
Gemini 3.5 Flash Has Left One Surface. Record the Address.
Google removed Gemini 3.5 Flash from one Gemini Enterprise app region on 4 August, while its API retirement table shows no shutdown date. The lesson is operational: a model name is not a lifecycle record. You need the deployment address.
Read article
The EU Delayed the Hardest Part of Its AI Act. Read It as a Caution, Not a Reprieve.
The world's most ambitious AI law just moved its own hardest deadline. High-risk obligations slip from August 2026 to December 2027, now written into Regulation (EU) 2026/1744. The delay is not a climbdown to celebrate. It is an admission the timeline was never workable, and a caution for anyone tempted to copy the EU model wholesale.
Read article
The EU Can Enforce Model Duties From 2 August. Your Vendor File Should Show It.
Europe's general-purpose AI rules did not suddenly begin in August. The sharper change is enforcement. Australian financial-services teams with EU exposure now need a version-specific evidence receipt that separates the vendor's duties from their own role and use.
Read article
Every Model You Rely On Has a Retirement Date
Claude Opus 4.1 stops answering on 5 August 2026. OpenAI shut down sixteen model snapshots on 23 July 2026 and will retire its Agent Builder and Evals platform on 30 November 2026. These dates are published, they are not negotiable, and the notice you receive depends on a product tier most teams never consciously chose.
Read article
Agents That Can Pay: Gate the Wallet First
The rails for agents that pay are being laid in public. Google's Agent Payments Protocol, built with more than sixty payment and technology firms, turns an agent's spending authority into a signed, limited, auditable instruction. Before any tool with a wallet reaches an Australian workplace, the practical question is the action gate: which payments an agent may start, at what limit, with whose approval, and what gets logged.
Read article
Stop Detecting AI. Start Checking Provenance.
Asking whether a file was made by AI is the wrong question, because the answer is a guess. Provenance asks a different one: what does the file itself carry about where it came from, signed by someone who can be identified. The C2PA specification reached version 2.4 in April 2026, added a machine-readable AI disclosure assertion, and now embeds into HTML and structured text, not just photographs. It is a real control with three real limits.
Read article
Fine-Tuning Writes Your Data Into the Model
Every organisation customising AI faces the same build decision: prompt it, retrieve for it, or fine-tune it. It is usually argued on accuracy and cost. The axis nobody raises is reversibility. Retrieval leaves your data in a store you own, audit and delete. Fine-tuning copies it into a model artefact you cannot meaningfully un-write, which makes the customisation call a records decision before it is an engineering one.
Read article
Internal Audit's AI Say-Do Gap Now Has Numbers
Two 2026 surveys put hard numbers on something internal audit has felt for a year: adoption is racing ahead of capability. Gartner has 83 per cent of audit functions piloting or using AI, while only a small share feel able to embed it. The IIA and AuditBoard have 85 per cent alert to AI-enabled fraud and fewer than 40 per cent ready to detect it. For an Australian function, the gap is a benchmark to self-assess against, not a headline to read past.
Read article
Your AI Logs the Tokens. Not the Decision.
Your AI telemetry records what the call cost, not what the model was told or what it said. Every content attribute in the OpenTelemetry GenAI conventions ships opt-in, and the conventions warn the message content is likely to hold personal information. That is a privacy default, not an audit default, and the two pull in opposite directions. Here is the capture, redaction and retention decision nobody is being asked to make.
Read article
Structured Outputs: Make Your AI Return Data You Can Audit
Most teams still treat an AI answer as prose to read and re-key. Structured outputs change that. You define a schema, the model is forced to fill it, and you get validated, type-safe data your systems can check. That makes an AI pipeline auditable, but schema-valid is not the same as true, so the human verification step does not go away.
Read article
Google Missed Its Own Release Date, and That Is the Story
Google promised Gemini 3.5 Pro for June 2026. It is mid-July and the flagship still is not generally available, held back over quality Google will not sign off. In a market that worships launch velocity, a missed date is the underreported signal, and a live test of your own buying discipline.
Read article
A Real CVE in Your Agent-Building Tools
A critical flaw in Langflow, tracked as CVE-2026-33017, let anyone run code on exposed servers with a single unauthenticated request, and attackers used it to install cryptominers within a day of disclosure. The vulnerability was not in a model. It was in the low-code tool teams use to build agents. Here is what that means for anyone evaluating an agent-building platform.
Read article
Your AI Agent Can Remember Now. Govern What It Keeps.
In 2026 the major labs shipped persistent memory as a first-class agent feature, barely six weeks apart. An agent that remembers across sessions is a different thing to govern, and a new attack surface. Treat the memory store as a governed data asset with write rules, provenance, expiry and rollback, not invisible plumbing.
Read article
What 9,700 Real Users Actually Do With Claude
Most debate about AI and work runs on anecdote. Anthropic's June 2026 Economic Index, Cadences, offers something rarer: behavioural data on what roughly 9,700 real users do with AI, matched to how they feel about their careers. Almost every conversation produces real work, and the heaviest delegators are the most optimistic. Here is what a people leader should take from it.
Read article
Ground the Model, Do Not Trust Its Memory
A weak model with the right tool beat a strong model working from memory. Anthropic's biology-agent benchmark put numbers on it: accuracy ran as low as 16.9% from recall and cleared 99.7% once grounded in the authoritative source. The lesson holds for every regulated workflow that turns on a rule, rate or standard. Ground the model, do not trust its memory.
Read article
Canberra Just Measured What AI Is Doing to Jobs. It Found 2%.
The Australian Government has published its first systematic measurement of AI's effect on employment. DEWR's July 2026 report finds the most AI-exposed occupations grew 5.6 per cent since ChatGPT arrived, against 9.5 per cent for the least exposed, and models a 2 per cent shortfall against trend. The number is small, the caveats are real, and the implications for WHS, workers compensation, GRC, HR and leadership teams start now.
Read article
Your AI Passed the Pilot. Production Is a Different Test.
A pilot that passed proves the AI worked once, on that data, on that version of the model. Then the model changes underneath you, the data shifts and usage drifts, and the thing you signed off quietly stops being the thing running. The answer is a standing evaluation harness, not a launch-day test. Here is how to build one, and why regulated work needs it most.
Read article
More Agents Is Not More Intelligence. Govern the Coordination.
The plumbing for multi-agent AI just got standardised, so building systems where agents talk to each other is now easy. That is exactly why the useful question changed. Not can you wire agents together, but should you, and how do you govern the coordination when more agents is not more intelligence.
Read article
GPT-Live Makes Voice a Work Interface. Check What Gets Recorded
OpenAI's GPT-Live can listen and speak at the same time, making voice feel closer to a continuous work interface than a turn-by-turn chatbot. It is rolling out to consumer ChatGPT plans first, not Business, Enterprise or Edu. That gap makes recording, retention and disclosure rules the immediate workplace question.
Read article
AI's Next Constraint Is Power. Australia Has Started Writing the Rules
Australia is starting to treat AI infrastructure as an energy-system issue. Government expectations now ask new data centres to add clean supply, pay their grid costs and support flexible demand, while a draft AEMC rule would set technical connection standards for large loads.
Read article
Model Routing Cuts AI Bills. It Also Moves Your Data.
The enterprise AI story has shifted from which model is best to which one you can afford to keep using, and buyers are moving from tokenmaxxing to model routing. The instinct is right. But for Australian regulated work a cheaper model is usually a different provider in a different place, so every routing rule is also a Privacy Act and APRA data-flow decision.
Read article
Someone Poisoned the Tool Description. The Agent Did the Rest.
Microsoft's incident-response team has documented the first real-world attack on the Model Context Protocol: poison a tool's description and you redirect the agent, without touching its code, its credentials or its prompt. The site's earlier MCP piece predicted this. Here is what "least agency, not just least privilege" means as an operational control, not a slogan.
Read article
A Government Now Vets Who Gets the Model. File It as a Vendor Risk.
For the second time in a month, a US government put a frontier model behind a gate. This time it was an access list: roughly 20 vetted organisations get GPT-5.6, chosen name by name. The capability story is covered. The one that lands on your desk is procurement: government-imposed access conditions are now a vendor-risk category your due diligence has to name.
Read article
AI Guardrails: The Safety Layer No Vendor Can Ship for You
A new default model does not make your AI system safe. Reliability and safety in production come from the guardrail layer you build around the model: input rails, grounding, output rails and action gates. Here is the architecture, a worked example that turns a written policy into running rails, and the prompts and checklist to build your first rail this week.
Read article
You Are Now Buying Software for Agents, Not People
Gartner says $234 billion of enterprise SaaS spend is exposed to agentic arbitrage by 2030. Read as a vendor problem it is a headline. Read as a buyer it changes how you procure and govern the software you already run, so this piece turns the forecast into an API-parity test you can run this week, a five-clause contract checklist and a Monday workflow.
Read article
Claude Sonnet 5 Became the Default. That Is a Change Event.
Anthropic released Claude Sonnet 5 on 30 June and made it the default in Claude Code and on Claude.ai Free and Pro. Almost nobody chooses a model, so a frontier swap is a silent change to a system you may have already validated. Here is how to treat it as a change event: the prompts, the Monday workflow and the register entry.
Read article
Fable 5 Returns With a Jailbreak Severity Framework
Claude Fable 5 comes back globally on 1 July, three weeks after a US directive pulled it. The return matters less than what came with it: a safety patch that over-blocks routine coding, and a four-dimension jailbreak-severity framework the major labs are building together.
Read article
AI Cyber Defence Just Scaled Up. Mind Your Open-Source Dependencies.
OpenAI pointed its most capable cyber model at the open-source software the world runs on, and found hundreds of real flaws in days. The capability is dual-use. Here is what it means for Australian teams.
Read article
AI Agents Just Went From Minutes to Hours. The Control Point Is Where They Run.
OpenAI bought Ona and published research showing AI agents now run for hours, not seconds. The unit you have to govern moved from the prompt to the environment the agent runs in.
Read article
GPT-5.6 Sol Lands. The Frontier Just Got Gated.
On 26 June OpenAI previewed GPT-5.6, a new three-model family it calls its strongest yet. The capability is real, but the news is the access: at the US government's request it launched to a handful of vetted partners, not to you. For the second time in a month a US lab's best model is gated by government. Frontier access is now a policy variable, and your AI plan needs to assume it.
Read article
Context Engineering: What the Model Is Allowed to See
The reliability of an AI system is decided less by how you word the prompt and more by what you let into the context window. Context engineering is the named discipline for that, and it is the highest-leverage AI skill for anyone doing real work, especially in regulated settings.
Read article
OpenAI Built Its Own Chip. The Real Story Is the Cost of Intelligence.
On 24 June, OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom chip, built to run its models more cheaply. The coverage is about a strike at Nvidia. The useful signal for a practitioner is the opposite of hardware: it is the falling cost of intelligence, and what that does to your AI decisions and your governance. Cost has quietly been holding AI back. That fence is coming down.
Read article
Your AI Assistant Just Became a Shared Teammate. Govern the Channel.
On 23 June, Anthropic launched Claude Tag, a single shared Claude that lives in a Slack workspace with its own memory and admin-scoped access to channels, tools and data. The unit of AI collaboration just moved from the private conversation to the team channel. The thing you now have to govern is no longer a prompt. It is a standing presence. Here is what changes, and the three decisions to make before it is live.
Read article
Model Context Protocol: The Standard Wiring AI Into Your Tools
In eighteen months the Model Context Protocol went from an Anthropic experiment to the way AI plugs into your tools and data. Understanding what it is matters less than governing the connectors, because each one is a new door into your systems. Here is the capability and the control work.
Read article
The Strongest Open Model Is Now Chinese. Mind Where Your Data Goes.
On 16 June, China's Z.ai released GLM-5.2 under an MIT licence with no regional limits, the highest-ranked open-weights model on its own coding benchmarks. With Anthropic's Fable 5 pulled by a US directive, the strongest model you can simply download and run is now Chinese. The decision that carries your risk is not the model. It is whether you run the open weights yourself or send your data to the hosted API.
Read article
ChatGPT Just Got Better at Health. Mind the Boundary.
On 18 June OpenAI announced a substantial step up in ChatGPT's health intelligence, free to the 230 million people who already ask it health questions every week. Better answers do not move the boundary between information and a clinical decision. Here is what that means for Australian professionals this week.
Read articleNews Posts
Rapid updates when AI news breaks
Concise rapid updates when AI news breaks. 150-200 words, no filler, straight to the signal. Available on the site and cross-posted to X (@TheAICommand) and Instagram (@the_aicommand).
Build 2026 makes agents first-class citizens of the Microsoft stack
OAIC survey finds trust in AI companies has collapsed to 4 per cent
Mistral puts frontier-class weights on four GPUs with Medium 3.5
ASIC research maps AI spreading through underwriting and claims
Google I/O 2026: agents get desktops, sandboxes and enterprise plumbing
ASIC demands urgent cyber uplift as frontier AI raises the threat level
Interactive tools